Document parsing for RAG: what to look for in a converter
When choosing a document parser for a RAG pipeline, judge it on what it does when a file cannot be read - a parser that silently indexes an empty page costs more than one that refuses it loudly.
Every parser handles the easy formats
Word, PowerPoint, Excel, Markdown, HTML and text-layer PDFs all extract reliably almost everywhere, because the text is stored as text. If your corpus is these formats, extraction is close to a solved problem and you should be choosing on cost, privacy and speed rather than on quality.
The differences live at the edges
Scanned PDFs, photographs of documents, screenshots, audio, video, merged table cells and right-to-left text are where parsers actually diverge. So does the boring case of a folder holding two hundred files where three are corrupt: a parser that aborts the batch is unusable, and one that drops them silently is worse.
Ask what happens on failure
The property worth testing is what lands in your index when a file cannot be read. Three outcomes exist: the file is skipped and reported, the file is skipped silently, or a placeholder is indexed. Only the first is safe. The third is actively harmful, because those placeholder passages resemble each other and outrank real content.
A manifest is not a nice-to-have
You cannot audit an index you cannot describe. A conversion should tell you how many files it read, which it skipped and why, how many chunks each produced, what embedded them, and whether any text was transcribed rather than extracted - OCR output is a machine's best reading of pixels and can be confidently wrong in ways parsed text cannot. RAG Converter writes all of that to manifest.json beside the chunks.
Where the parsing happens is a real choice
Uploading a corpus to a hosted parser means the documents leave your control, which for legal, medical or commercially sensitive files is the whole decision. The browser can do more of this than people assume: RAG Converter parses documents and reads images in the tab, with nothing uploaded, and offers the server only for the work a browser genuinely cannot do.